著者

若林 啓 竹内 誉羽 平松 淳 中野 幹生

出版者

一般社団法人 人工知能学会

巻号頁・発行日

pp.1N3J903, 2019 (Released:2019-06-01)

---

本論文では,音声言語理解の代表的な形式であるスロットフィリングタスクを,入力文の最適分割を求める問題として定式化するアプローチを提案する.この定式化を用いることで,系列ラベリングとして定式化して深層学習を適用する従来の手法に比べ,小さい計算資源で実行可能な言語モデルの学習に基づいた手法を導くことができる.提案手法は,モデルの学習をワンパスアルゴリズムによって高速に行い,最も確率の高い解釈の推定を動的計画法を用いて効率的に行う.実験により,提案手法は深層学習手法と同等の推定精度を達成しつつ,高速かつ省メモリに動作することを確認する.

著者

竹内 誉羽 庄野 修 辻野 広司

出版者

一般社団法人 人工知能学会

雑誌

人工知能学会論文誌 (ISSN:13460714)

巻号頁・発行日

vol.27, no.2, pp.92-102, 2012 (Released:2012-02-10)

参考文献数

23

---

Future robots/agents will perform situated behaviors for each user. Flexible behavioral learning is required for coping with diverse and unexpected users' situations. Unexpected situations are usually not tractable for machine learning systems that are designed for pre-defined problems. In order to realize such a flexible learning system, we were trying to create a learning model that can function in several different kinds of state transitions without specific adjustments for each transition as a first step. We constructed a modular neural network model based on reinforcement learning. We expected that combining a modular architecture with neural networks could accelerate the learning speed of neural networks. The inputs of our neural network model always include not only observed states but also memory information for any transition. In pure Markov decision processes, memory information is not necessary, rather it can lead to lower performance. On the other hand, partially observable conditions require memory information to select proper actions. We demonstrated that the new learning model could actually learn those multiple kinds of state transitions with the same architectures and parameters, and without pre-designed models of environments. This paper describes the performances of constructed models using probabilistically fluctuated Markov decision processes including partially observable conditions. In the test transitions, the observed state probabilistically fluctuated. The new learning model could function in those complex transitions. In addition, the learning speeds of our model are comparable to a reinforcement learning algorithm implemented with a pre-defined and optimized table-representation of states.